New models · Compared
GPT-5.6 Sol vs GPT-6 Sol: Worth the Switch?
GPT-6 Sol replaced GPT-5.6 Sol as OpenAI's mid-priced model on September 22, 2026, at half the API price: $2 and $10 per million input and output tokens, down from $4 and $20. OpenAI also says it makes about half as many factual mistakes as its predecessor.
Both models were still in ChatGPT Work's list on September 23, so we gave them the same small-business brief at the same Medium effort and compared what each put in front of customers.
What changed on paper
OpenAI trained GPT-6 Sol with methods similar to GPT-6 Astra and cut the price in half. Its launch post says that on an internal factuality test built from conversations where users had flagged errors, GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol. That is OpenAI's measurement, not ours.
Both models answer in text only and share the same 1.05 million-token context window.
| Model | Input | Output |
|---|---|---|
| GPT-5.6 Sol (promotional price) | $4 | $20 |
| GPT-6 Sol | $2 | $10 |
Our test
Both models got the Kettle & Crumb brief: a fictional bakery launching a $9 Pumpkin Rye Sourdough on Saturday 3 October, 60 loaves, preorders closing Thursday, no discount, no delivery. The brief tells the model to list anything it guesses. Then we sent a follow-up that asks for things the facts do not contain: a usual Friday offer and opening hours.
Now write a second, shorter caption for Friday 2 October. Mention our usual Friday offer and our opening hours.The difference that matters
Both models got every fact right, and both refused to invent the Friday offer or the opening hours, leaving placeholders instead. Timing was close: 15 seconds for GPT-6 Sol and 17 seconds for GPT-5.6 Sol.
The difference was one line. GPT-5.6 Sol's caption told customers to pick up from 14 Alder Street. The facts only say there is no delivery; pickup is a reasonable guess, but it went into the public caption without appearing in the assumptions list. GPT-6 Sol put no guess in either caption.
How GPT-6 Sol wrote the brief
GPT-6 Sol was the most literal model we tested. Its prompts described one loaf on a plain surface, and it listed the surface, light and camera angles as its own choices. That matches the better factuality OpenAI reports — and it also means you should supply the look of your product yourself.
Should you switch?
For creative planning, yes: GPT-6 Sol is half the price and, in our one run, kept guesses out of the customer-facing text. Keep the brief habit that made the test fair — list the facts, ask for assumptions — because a model that guesses less still guesses.
Ciyo runs neither Sol model. Its agent offers Ciyo Agent and GPT-6 Astra, and either Sol's prompts can be pasted into the Image Generator.
An independent price check
Benchmark trackers measured the same price move on launch day.
Posted 2026-09-22 by Artificial Analysis, an independent benchmark organisation. It reports that Sol and Luna halve cost relative to GPT-5.6 Sol and Luna on its index. That is their measurement of general tasks, not of creative briefs.
GPT-5.6 Sol vs GPT-6 Sol
Is GPT-6 Sol cheaper than GPT-5.6 Sol?
Yes. $2 and $10 per million input and output tokens, against $4 and $20 for GPT-5.6 Sol's promotional price.
Is GPT-6 Sol more accurate?
OpenAI reports about half as many factual mistakes. In our one brief, GPT-6 Sol put no guess in its captions; GPT-5.6 Sol put one.
Is GPT-5.6 Sol still available?
It was still in ChatGPT Work's model list on September 23, 2026.
Does Ciyo use either Sol model?
No. Ciyo's agent offers Ciyo Agent and GPT-6 Astra.
Keep the facts in, make the picture in Ciyo
Plan with whichever model you trust, then generate and compare images on one Ciyo board.